Diagnosis of pleural effusion and pneumothorax

Chest x-ray images from the MIMIC-CXR database were used for two diagnostic tasks: 17,102 images were used for the diagnosis of pleural effusion and 15,838 images were used for the diagnosis of pneumothorax. The dataset provides ground truth diagnoses for the two conditions and metadata including view positions¹⁴. For both diagnostic tasks, patients were divided into non-overlapping positive and negative groups depending on whether they had at least one positive chest x-ray for the condition of interest. Then, anteroposterior (AP) view images were selected for every patient and the majority class was down sampled. The process resulted in 17,102 images for the pleural effusion dataset, and 15,838 images for the pneumothorax dataset.

For all diagnostic tasks, the data was split into training and testing datasets based on a 70–30 ratio.

All analyses were performed in accordance with the guidelines and regulations stated in the PhysioNet Credentialed Health Data License: https://physionet.org/content/mimiciii/view-license/1.4/ and https://physionet.org/content/mimic-cxr/view-license/1.0.0/.